Research Internships

Classification of Occluded Objects by Neural Networks

Summer project done with Prof. Gabriel Kreiman, Boston Children's Hospital, Harvard Medical School.

Convolutional Neural Networks (CNNs) have been known to do a great job in classifying diverse varieties of images. Their architecture is based upon the feed-forward information transfer of the ventral visual pathway. But the primate visual system is not only good at classifying whole objects; it excels are perceiving and classifying occluded objects as well. Deep Neural Networks, in spite of displaying remarkable similarities with the information processing and transfer systems of the primate visual pathway does not seem to be as good at the task of classifying occluded objects, as human beings. We suspected that introducing connections in the CNNs that make it more similar to the neural connections known in the ventral visual pathway might make them better at the task of classifying occluded objects, than the standard feed-forward networks. To be more precise, we predicted that introducing local within-layer connections between neurons that display similar preference (based upon the observation that neurons in the cortex with similar preference are more connected to each other) would improve their performance at pattern completion, and thus, improve their performance at the task of classifying occluded objects.

Cellfiefuge: Auto-inducing and Auto-aggregating Bacteria

Project done as part of IISc iGEM Team, 2016, for International Genetic Engineering Competition (iGEM), organised by the iGEM Foundation, at Boston, USA.

With the advent of rDNA technology in the late 1970s, medicine, agriculture, and several other areas underwent a quantum leap and from that point, progress only hastened, from one only one recombinant pharmaceutical approved for human use (insulin) in 1982 to one hundred and fifty-one FDA approved protein based recombinant pharmaceuticals by 2009. Despite being in high demand recombinant products are expensive due to several factors like long and expensive development time, a high failure rate of the products developed, manufacturing costs requiring expensive technologies and processes and the involvement of skilled labor on both the manufacturing and the healthcare provider’s side. As scientists and engineers, it seems obvious that our contribution can be most easily and effectively be made at the level of manufacturing costs; to try to bring down the cost of these life-saving products. To reduce the cost, we wanted to optimize recombinant protein production in bacterial systems, and thus designed and introduced a genetic circuit in them which allowed them to start protein production at an optimal time. To implement this, we used the previously known principle of quorum sensing and used it as a switch to begin protein production, when the bacterial population reached an optimal concentration. We added a second switch in the genetic circuit, inspired by the principles of diauxic growth, that made the bacteria produce a secondary surface adhesion protein, which in turn made the bacteria clump and precipitate, enabling them to be harvested without the capital-heavy process of centrifugation.

McFarland Standardization in M. smegmatis

Summer project done with Prof. Dipankar Chatterji, Molecular Biophysics Unit, Indian Institute of Science.

Diversity is one of the key properties biological systems display, that makes them extremely interesting to study. But the same property also makes the task of studying them more difficult. A metric developed for quantification of a particular set of organisms might not hold for another set, even if both the sets share a lot of common properties. For more than a century, McFarland standards have been used to estimate the concentration of bacterial cultures in terms of colony forming units (CFUs). There is a linear correlation between McFarland standards and CFUs which holds very accurately for enterobacterial species, and these standards are also used extensively for estimating mycobacterial CFUs. However, we did not come across any report that has recorded relationship between McFarland standards and mycobacterial CFUs. In this work, we have determined the existence of a linear correlation between mycobacterial CFUs and McFarland standards using Mycobacterium smegmatis as a model system. Surprisingly, we also found that for a given McFarland standard, mycobacterial CFUs are two orders of magnitude less compared with enterobacterial CFUs. This finding has implications in clinical settings where McFarland standards are used to estimate minimum inhibitory concentrations (MICs) for antimicrobials for mycobacteria.

Protein Designing and Simulated Optimization of Protein Cores

Voluntary project done with Prof. Nagasuma Chandra, Department of Biochemistry, Indian Institute of Science.

It is known that the tertiary structure of a protein molecule can be predicted with high confidence from its primary structures. The reverse mapping, on the other hand, is not unique - very similar tertiary structure can be achieved by very different amino-acid sequences. Hence, it is possible that the sequence of a protein can be radically changed while maintaining its structure, and end up with a sequence that makes the protein more stable by minimizing its energy. We worked with leucine-rich protein cores as a model and tried to optimize its sequence, using the simulated annealing algorithm. We introduced random mutations in the protein and tried to observe which mutations caused a reduction of the potential energy of the hydrophobic core and whether we can find a sequence that would minimize this energy.

We also worked on a secondary, independent, project that tried to address the problem of finding a reliable probe for the detection of tuberculosis. Antigen-85 (Ag85) is a well-known tuberculosis marker, but it has no known probe that is small in size and binds with it with high efficacy and specificity. Using an Immunoglobulin-M (IgM) binding scaffold as a backbone, we computationally designing a protein and replaced a Type-II Beta Hairpin in it, such that it can bind to Ag85. The newly designed protein was attached with a green fluorescent protein (GFP) tag, for easier localization and isolation of the protein.